Float, Half and BFloat16 mismatch: trace the failing operation

Symptoms and scope

A dtype failure differs from a shape or device failure. Start with the traceback and incompatible operand types; a KSampler error alone does not identify the component that caused them.

Error fragments for search; IDs, values and filenames may vary:

mat1 and mat2 must have the same dtype
Float and Half
BFloat16

Source-supported context

Issue #6907 records a Float/Half failure after a Mac Desktop update. Its traceback reaches an attention torch.baddbmm call through several third-party extension frames; the report does not prove which component introduced the mismatch. BFloat16 is a general search term here, not an error reported in that case. PyTorch AMP guidance treats a type mismatch as a reason to inspect the operation, not to insert blanket casts. Source 1 Source 2 Source 3

Cases to distinguish

1. A loader or custom node casts only part of the computation path.

2. Quantization, attention or device fallback changes the compute dtype.

3. Partially updated core/extensions retain incompatible precision assumptions or startup options.

Diagnostic sequence

The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.

Step 1. Keep the exact dtype message, backend and full traceback. Find the failing operation and any third-party frames; do not assign blame from the KSampler node label alone. Separate stored weight precision from runtime tensor dtype.

Step 2. In a copy, compare the model or workflow author's documented backend and precision settings when available. Temporarily remove only optional quantization or acceleration, one change at a time, and monitor memory.

Step 3. Compare a same-family minimal native workflow, then reintroduce LoRA, controls and extensions one at a time.

Step 4. Report a stable minimal reproduction with versions. Source-level forced casting belongs to someone who understands the implementation, not a generic one-click patch.

Completion check

The operation completes with compatible types and valid output without a new memory or numerical failure. A resulting OOM is a separate problem.

Limitations and cautions

Converting everything to float32 can substantially increase memory; converting everything to half can introduce unsupported operations or numerical instability.

Original sources

• Float and Half mismatch on Mac Desktop — checked 2026-09-26.
• PyTorch AMP recipe — original record checked 2026-09-26.
• ComfyUI custom node troubleshooting — original record checked 2026-09-26.

English source review: 2026-09-26. No site-run workflow, GPU test, minimum-memory measurement or repair outcome is claimed.

Related troubleshooting and guides

• Expected all tensors to be on the same device: inspect branch boundaries
• Float8 is unsupported on MPS: distinguish stored weights from compute dtype
• mat1 and mat2 shapes cannot be multiplied: inspect the model component chain